RSIAgent, a new research framework posted on Hugging Face, proposes a different route to improving digital agents: instead of scaling model parameters, it scales experience. The system is described as training-free, meaning the underlying model weights stay frozen throughout the improvement process.
The framework coordinates three types of agents — a curriculum agent, an actor agent, and a verifier agent — that continually explore an environment, validate outcomes, and retain environment-specific knowledge. According to the authors, this memory includes reusable causal relationships linking actions, conditions, and consequences.
RSIAgent uses a broad-then-deep exploration strategy. Broad, parallel recursive self-exploration is meant to discover diverse environment structures, while focused deep exploration targets hard cases, hidden constraints, boundary conditions, and previously unknown causal dependencies. The resulting memory is frozen and can be reused for downstream tasks without updating model parameters.
On OSWorld-v2 and Agent's Last Exam, the authors report that RSIAgent substantially improves strong open-source models, enabling Kimi-K3 and GLM-5.3 to outperform frontier closed-source models including GPT-6. The team frames these results as evidence that agents can keep improving by autonomously acquiring, verifying, and reusing their own experience.
Code, a blog post, a project page, and the paper are linked from the announcement, and the authors say they welcome feedback.